Automatic Accessment of Phenotypes in lettuce plants by using Chlorophyl Fluorescence Kinetics and Machine Learning

Agriculture aims at increasing production and provision of high quality products to the market. Most of the times, quality is strongly correlated the variety or the hybrid. Specifically, lettuce qualitative characteristics and nutrients appear to vary strongly in different varieties and hybrids. In the current research, lettuce plants were harvested at baby, immature and mature stage in 46, 60 and 70 days of growth, respectively. Then, the parameters of chlorophyll fluorescence were determined in two middle leaves of 3 plants of each hybrid at different harvesting stages by using chlorophyll fluorescence kinetics. The measurements revealed significant differences between varieties and hybrids. The fluorescence parameters were utilized as inputs for training different models of different novelty detection methods aiming at the identification of the phenotype of different varieties and hybrids. It is already known that novelty detection can be easily combined with machine learning techniques so as to detect abnormal events. This system was capable of diagnosing a new fault that did not appear in the training data set. For change detection, a normality description (baseline condition) was constructed. As a result, deviations were detected from this description of the normal domain (as new varieties and hybrids). In this paper, an active learning method based on novelty detection in the form of one-class classifiers for the iterative detection of different phenotypes of lettuce plants based on fluorescence sensing is proposed. This method learns to distinguish between different hybrids of lettuce plants based on their fluorescence parameter differences. The proposed active learning method uses one-class classification to detect phenotypic deviations between hybrids as outliers which are then augmented in a baseline multiclass classifier which is subsequently ready for novelty detection in case new phenotypes appear. Then, the procedure continues for the next hybrid and the proposed scheme can learn and augment new phenotypes indefinitely. It was shown that the identification of different lettuce phenotypes corresponding to hybrids is precise by the proposed Active Learning Method due to non-linearity problem which is due to the heterogeneity of the fluorescence kinetics parameters.